In most cities in China, surveillance video is the main means of collecting traffic information. The video method collects traffic volume, which has the advantages of saving manpower and processing large amount of information. However, it is impossible to carry out detailed division of different vehicle objects like manual survey, despite video capture. It has been able to overcome the interference of the environment and accurately count the vehicles, but it is not possible to convert different vehicles, resulting in the final traffic volume is often lower than the actual production, causing the relevant design management personnel to make a wrong decision. To this end, this paper expects to realize the conversion of video traffic volume by using the method of transfer learning.

CICTP